[X86][SSE] Improve combineLogicBlendIntoPBLENDV to use general masks.

[RKSimon]

Summary

Currently combineLogicBlendIntoPBLENDV can only match ASHR to detect sign splatting of a bit mask, this patch generalises this to use computeNumSignBits instead.

This is a first step in several things we can do to improve PBLENDV support:

• Better matching of X86ISD::ANDNP patterns.
• Handle floating point cases.
• Better vector and bitcast support in computeNumSignBits.
• Recognise that PBLENDV only uses the sign bit of the mask, we should be able strip away sign splats (ASHR, PCMPGT isNeg tests etc.).

[delena]

lib/Target/X86/X86ISelLowering.cpp

31566

Why VT should be v2i64 of v4i64? Isn't the same transformation profitable for other integer types?

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

31566

It's due to us promoting to v2i64/v4i64 for X86ISD::ANDNP - it's not actually necessary but seems to have been used as an early out.

I'll remove it - it won't make any difference to the codegen now but will make it easier to perform more thorough ANDNP matching mentioned in the TODO.

[RKSimon]

Addressed Elena's comments.

[delena]

test/CodeGen/X86/pr32907.ll

26

I assume that ComputeNumSignBits() does not recognize the ASHR sequence here, otherwise it would be able to generate VPBLEND, right?

[RKSimon]

test/CodeGen/X86/pr32907.ll

26

Exactly - this is an example of ones of the TODOs - the ANDNP is being generated at the same time that the ASHR_v2i64 is being lowered into the PSRAD+PSHUFD. If we could recognise the AND(XOR(-1,M), X) pattern earlier it would combine.

Note it wouldn't generate a VPBLEND, it would generate the SUB(XOR(X, M), M) pattern similar to the AVX512 codegen.

[spatel]

test/CodeGen/X86/vselect-pcmp.ll

144

You can remove those snarky 16-bit comments. 😃

154–156

Looks like we have to go to extremes to get the AVX1 case although this may improve with a patch that I'm working on for PR32790:
https://bugs.llvm.org/show_bug.cgi?id=32790

lib/Target/X86/X86ISelLowering.cpp

 static SDValue combineLogicBlendIntoPBLENDV(SDNode *N, SelectionDAG &DAG,
                                             const X86Subtarget &Subtarget) {
   assert(N->getOpcode() == ISD::OR);
 
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
   EVT VT = N->getValueType(0);
 
   if (!((VT == MVT::v2i64) || (VT == MVT::v4i64 && Subtarget.hasInt256())))
     return SDValue();
   assert(Subtarget.hasSSE2() && "Unexpected i64 vector without SSE2!");
 
-  // Canonicalize pandn to RHS
-  if (N0.getOpcode() == X86ISD::ANDNP)
+  // Canonicalize AND to LHS.
+  if (N1.getOpcode() == ISD::AND)
     std::swap(N0, N1);
 
+  // TODO: Attempt to match against AND(XOR(-1,X),Y) as well, waiting for
+  // ANDNP combine allows other combines to happen that prevent matching.
   if (N0.getOpcode() != ISD::AND || N1.getOpcode() != X86ISD::ANDNP)
     return SDValue();
 
   SDValue Mask = N1.getOperand(0);
   SDValue X = N1.getOperand(1);
   SDValue Y;
   if (N0.getOperand(0) == Mask)
     Y = N0.getOperand(1);
   if (N0.getOperand(1) == Mask)
     Y = N0.getOperand(0);
 
   // Check to see if the mask appeared in both the AND and ANDNP.
   if (!Y.getNode())
     return SDValue();
 
   // Validate that X, Y, and Mask are bitcasts, and see through them.
   Mask = peekThroughBitcasts(Mask);
   X = peekThroughBitcasts(X);
   Y = peekThroughBitcasts(Y);
 
   EVT MaskVT = Mask.getValueType();
-
-  // Validate that the Mask operand is a vector sra node.
-  // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
-  // there is no psrai.b
   unsigned EltBits = MaskVT.getScalarSizeInBits();
-  unsigned SraAmt = ~0;
-  if (Mask.getOpcode() == ISD::SRA) {
-    if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
-      if (auto *AmtConst = AmtBV->getConstantSplatNode())
-        SraAmt = AmtConst->getZExtValue();
-  } else if (Mask.getOpcode() == X86ISD::VSRAI)
-    SraAmt = Mask.getConstantOperandVal(1);
 
-  if ((SraAmt + 1) != EltBits)
+  // TODO: Attempt to handle floating point cases as well?
+  if (!MaskVT.isInteger() || DAG.ComputeNumSignBits(Mask) != EltBits)
     return SDValue();
 
   SDLoc DL(N);
 
   // Try to match:
   //   (or (and (M, (sub 0, X)), (pandn M, X)))
   // which is a special case of vselect:
   //   (vselect M, (sub 0, X), X)
auto IsNegV = [](SDNode *N, SDValue V) {
         ISD::isBuildVectorAllZeros(N->getOperand(0).getNode());
     };
     SDValue V;
     if (IsNegV(Y.getNode(), X))
       V = X;
     else if (IsNegV(X.getNode(), Y))
       V = Y;
 
-    if (V) {
-      if (EltBits != 8 && EltBits != 16 && EltBits != 32)
-        return SDValue();
-
+    if (V && DAG.getTargetLoweringInfo().isOperationLegal(ISD::SUB, VT)) {
       SDValue SubOp1 = DAG.getNode(ISD::XOR, DL, MaskVT, V, Mask);
       SDValue SubOp2 = Mask;
 
       // If the negate was on the false side of the select, then
       // the operands of the SUB need to be swapped. PR 27251.
       // This is because the pattern being matched above is
       // (vselect M, (sub (0, X), X)  -> (sub (xor X, M), M)
       // but if the pattern matched was
       // (vselect M, X, (sub (0, X))), that is really negation of the pattern
       // above, -(vselect M, (sub 0, X), X), and therefore the replacement
       // pattern also needs to be a negation of the replacement pattern above.
       // And -(sub X, Y) is just sub (Y, X), so swapping the operands of the
       // sub accomplishes the negation of the replacement pattern.
       if (V == Y)
          std::swap(SubOp1, SubOp2);
 
-      return DAG.getBitcast(VT,
-                            DAG.getNode(ISD::SUB, DL, MaskVT, SubOp1, SubOp2));
+      SDValue Res = DAG.getNode(ISD::SUB, DL, MaskVT, SubOp1, SubOp2);
+      return DAG.getBitcast(VT, Res);
     }
   }
 
   // PBLENDVB is only available on SSE 4.1.
   if (!Subtarget.hasSSE41())
     return SDValue();
 
   MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;

test/CodeGen/X86/cast-vsel.ll

 ; SSE2-NEXT:    pandn %xmm4, %xmm0
 ; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc:
 ; SSE41:       # BB#0:
 ; SSE41-NEXT:    pcmpeqw %xmm1, %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
-; SSE41-NEXT:    pshufb %xmm1, %xmm5
-; SSE41-NEXT:    pshufb %xmm1, %xmm4
-; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
 ; SSE41-NEXT:    pshufb %xmm1, %xmm3
 ; SSE41-NEXT:    pshufb %xmm1, %xmm2
 ; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; SSE41-NEXT:    pand %xmm0, %xmm2
-; SSE41-NEXT:    pandn %xmm4, %xmm0
-; SSE41-NEXT:    por %xmm2, %xmm0
+; SSE41-NEXT:    pshufb %xmm1, %xmm5
+; SSE41-NEXT:    pshufb %xmm1, %xmm4
+; SSE41-NEXT:    punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
+; SSE41-NEXT:    pblendvb %xmm0, %xmm2, %xmm4
+; SSE41-NEXT:    movdqa %xmm4, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: trunc:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
 ; AVX1-NEXT:    vpshufb %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm3[0],xmm1[0]
-; AVX1-NEXT:    vpandn %xmm1, %xmm0, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
-; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
 ; AVX1-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
-; AVX1-NEXT:    vpand %xmm0, %xmm2, %xmm0
-; AVX1-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm2
+; AVX1-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
+; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
+; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
+; AVX1-NEXT:    vpblendvb %xmm0, %xmm1, %xmm2, %xmm0
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: trunc:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31]
-; AVX2-NEXT:    vpshufb %ymm1, %ymm3, %ymm3
-; AVX2-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[0,2,2,3]
-; AVX2-NEXT:    vpandn %xmm3, %xmm0, %xmm3
-; AVX2-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
+; AVX2-NEXT:    vpshufb %ymm1, %ymm2, %ymm2
+; AVX2-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX2-NEXT:    vpshufb %ymm1, %ymm3, %ymm1
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]
-; AVX2-NEXT:    vpand %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    vpor %xmm3, %xmm0, %xmm0
+; AVX2-NEXT:    vpblendvb %xmm0, %xmm2, %xmm1, %xmm0
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
   %cmp = icmp eq <8 x i16> %a, %b
   %sel = select <8 x i1> %cmp, <8 x i32> %c, <8 x i32> %d
   %tr = trunc <8 x i32> %sel to <8 x i16>
   ret <8 x i16> %tr
 }
 

test/CodeGen/X86/pr32907.ll

 ; SSE-NEXT:    por %xmm2, %xmm1
 ; SSE-NEXT:    movdqa %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX2-LABEL: PR32907:
 ; AVX2:       # BB#0: # %entry
 ; AVX2-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpsrad $31, %xmm0, %xmm1
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
 ; AVX2-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX2-NEXT:    vpsubq %xmm0, %xmm2, %xmm2
 ; AVX2-NEXT:    vpandn %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    vpand %xmm2, %xmm1, %xmm1
 ; AVX2-NEXT:    vpor %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: PR32907:
 ; AVX512:       # BB#0: # %entry
 ; AVX512-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vpsraq $63, %zmm0, %zmm1
-; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; AVX512-NEXT:    vpsubq %xmm0, %xmm2, %xmm2
-; AVX512-NEXT:    vpandn %xmm0, %xmm1, %xmm0
-; AVX512-NEXT:    vpand %xmm2, %xmm1, %xmm1
-; AVX512-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpxor %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpsubq %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 entry:
   %sub13.i = sub <2 x i64> %astype.i, %astype6.i
   %x.lobit.i.i = ashr <2 x i64> %sub13.i, <i64 63, i64 63>
   %sub.i.i = sub <2 x i64> zeroinitializer, %sub13.i
   %0 = xor <2 x i64> %x.lobit.i.i, <i64 -1, i64 -1>
   %1 = and <2 x i64> %sub13.i, %0
   %2 = and <2 x i64> %x.lobit.i.i, %sub.i.i
   %cond.i.i = or <2 x i64> %1, %2
   ret <2 x i64> %cond.i.i
 }

test/CodeGen/X86/vselect-pcmp.ll

 
 ; Sorry 16-bit, you're not important enough to support?
 
 define <8 x i16> @signbit_sel_v8i16(<8 x i16> %x, <8 x i16> %y, <8 x i16> %mask) {
 ; AVX-LABEL: signbit_sel_v8i16:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX-NEXT:    vpcmpgtw %xmm2, %xmm3, %xmm2
-; AVX-NEXT:    vpandn %xmm1, %xmm2, %xmm1
-; AVX-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; AVX-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %tr = icmp slt <8 x i16> %mask, zeroinitializer
   %z = select <8 x i1> %tr, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %z
 }
 
 define <4 x i32> @signbit_sel_v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %mask) {
 ; AVX12F-LABEL: signbit_sel_v4i32:

 ; AVX512-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX512-NEXT:    retq
   %tr = icmp slt <32 x i8> %mask, zeroinitializer
   %z = select <32 x i1> %tr, <32 x i8> %x, <32 x i8> %y
   ret <32 x i8> %z
 }
 
 ; Sorry 16-bit, you'll never be important enough to support?
 
 define <16 x i16> @signbit_sel_v16i16(<16 x i16> %x, <16 x i16> %y, <16 x i16> %mask) {
 ; AVX1-LABEL: signbit_sel_v16i16:
 ; AVX1:       # BB#0:
 ; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
 ; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
 ; AVX1-NEXT:    vpcmpgtw %xmm3, %xmm4, %xmm3
 ; AVX1-NEXT:    vpcmpgtw %xmm2, %xmm4, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
 ; AVX1-NEXT:    vandnps %ymm1, %ymm2, %ymm1
 ; AVX1-NEXT:    vandps %ymm2, %ymm0, %ymm0
 ; AVX1-NEXT:    vorps %ymm1, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: signbit_sel_v16i16:
 ; AVX2:       # BB#0:
 ; AVX2-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX2-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm2
-; AVX2-NEXT:    vpandn %ymm1, %ymm2, %ymm1
-; AVX2-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpor %ymm1, %ymm0, %ymm0
+; AVX2-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: signbit_sel_v16i16:
 ; AVX512:       # BB#0:
 ; AVX512-NEXT:    vpxor %ymm3, %ymm3, %ymm3
 ; AVX512-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm2
-; AVX512-NEXT:    vpandn %ymm1, %ymm2, %ymm1
-; AVX512-NEXT:    vpand %ymm2, %ymm0, %ymm0
-; AVX512-NEXT:    vpor %ymm1, %ymm0, %ymm0
+; AVX512-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
 ; AVX512-NEXT:    retq
   %tr = icmp slt <16 x i16> %mask, zeroinitializer
   %z = select <16 x i1> %tr, <16 x i16> %x, <16 x i16> %y
   ret <16 x i16> %z
 }
 
 define <8 x i32> @signbit_sel_v8i32(<8 x i32> %x, <8 x i32> %y, <8 x i32> %mask) {
 ; AVX12-LABEL: signbit_sel_v8i32: